Online deliberation platforms need a way to tell which ideas bridge a divided population, and the talk gives one that works from approve and disapprove votes alone; one of its measures is already in use in Change.org's Civic Dialogues. It then models how an LLM synthesizes new ideas from existing ones, which yields algorithms that provably find ideas a whole population can get behind, under a condition that can be tested empirically. It closes on a pilot deployment on Change.org. Joint work with Blair, Cookson, de Raaij, Procaccia, Rubin-Toles, Si and Wang.

Nisarg Shah is an Associate Professor of Computer Science at the University of Toronto, a Research Lead for Ethics of AI at the Schwartz Reisman Institute, winner of the 2024 IJCAI Computers and Thought Award, and a co-developer of Spliddit.org.

• Approve and disapprove votes are sparse: most participants see a few ideas. How does the bridging measure behave when coverage is uneven, and who decides which ideas each voter sees?
• What is the empirically testable condition, and did it hold on the Change.org pilot?
• A synthesized idea that everyone approves may be one that commits to nothing. Does the model penalize vagueness, or is that left to the population's votes?

Nothing here covers computational social choice. Earlier talks in the series: Marcussen and Gong: Quality Control for Pattern Counting, Adaptivity in Quantum Learning and David Alvarez-Melis: Making Data Count.

Notes pending.